Avoiding Replicates in Biocatalysis Experiments: Machine Learning for Enzyme Cascade Optimization

The optimization of enzyme cascades is a complex and resource-demanding task due to the multitude of parameters and synergistic effects involved. Machine learning can support the identification of optimal reaction conditions, for example, in the case of Bayesian optimization (BO), by proposing new experiments based on Gaussian process regression (GPR) and expected improvement (EI). Here, we used BO to optimize the concentrations of the reaction components of an enzyme cascade. The productivity-cost-ratio was chosen as the optimization objective in order to achieve the highest possible productivity, which was normalized to the costs of the materials used to prevent convergence to ever-increasing enzyme concentrations. To reduce the experimental effort, contrary to common practice in biological experiments, we did not use replicates but instead relied on the algorithm’s proposed experiments and inherent uncertainty quantification. This approach balances parameter space exploration and exploitation, which is critical for the efficient and effective identification of optimal reaction conditions. At the optimized reaction conditions identified in our study, the productivity-cost ratio was doubled to 38.6 mmol L-1 h-1 €-1 compared to a reference experiment. The parameter optimization required only 52 experiments while being robust to outlying experimental results.